Convergence

Citation
Barro, Robert J., and Xavier Sala-i-Martin. 1992. Convergence. Journal of Political Economy
100(2): 223-251.

Published Version
doi:10.1086/261816

Permanent link
http://nrs.harvard.edu/urn-3:HUL.InstRepos:3451299

Terms of Use
This article was downloaded from Harvard University’s DASH repository, and is made available
under the terms and conditions applicable to Other Posted Material, as set forth at http://
nrs.harvard.edu/urn-3:HUL.InstRepos:dash.current.terms-of-use#LAA

Share Your Story

The Harvard community has made this article openly available.
Please share how this access benefits you. Submit a story .

Accessibility
Convergence

RobertJ. Barro
Harvard Universityand National Bureau of EconomicResearch

XavierSala-i-Martin
Yale Universityand National Bureau of EconomicResearch

A key economic issue is whether poor countries or regions tend to

grow faster than rich ones: are there automatic forces that lead to
convergence over time in the levels of per capita income and prod-
uct? We use the neoclassical growth model as a framework to study
convergence across the 48 contiguous U.S. states. We exploit data
on personal income since 1840 and on gross state product since
1963. The U.S. states provide clear evidence of convergence, but
the findings can be reconciled quantitatively with the neoclassical
model only if diminishing returns to capital set in very slowly. The
results for per capita gross domestic product from a broad sample
of countries are similar if we hold constant a set of variables that
proxy for differences in steady-state characteristics.

A key economic issue is whether poor countries or regions tend to

grow faster than rich ones: are there automatic forces that lead to
convergence over time in the levels of per capita income and product?
We use the neoclassical growth model as a framework to study con-
vergence across the 48 contiguous U.S. states. We exploit data on
personal income since 1840 and on gross state product since 1963.
For studying the determinants of economic growth, the experience

We have benefited from research support by the National Science Foundation,the

Bradley Foundation, and the Hoover Institution. We appreciate helpful comments
from Gary Becker, Paul David, Steve Durlauf, Susan Guthrie, Carol Heim, Anne
Krueger, Edward Lazear, Bob Lucas, Greg Mankiw, Kevin M. Murphy, Danny Quah,
Sergio Rebelo, and Gavin Wright.
[Journal of Politzcal Economy, 1992, vol. 100, no. 2]
? 1992 by The University of Chicago. All rights reserved. 0022-3808/92/0002-0001$01.50

223
224 JOURNAL OF POLITICAL ECONOMY

of the U.S. states represents a vastly underutilized resource: in effect,

we have over a century of data on 48 economies (although surely not
48 closed economies!).
The U.S. states provide clear evidence of convergence in the sense
that poor economies tend to grow faster than rich ones in per capita
terms. The estimated speed of convergence accords with the neoclas-
sical growth model if we take a broad view of capital so that diminish-
ing returns to capital set in slowly as an economy develops. The find-
ings for the U.S. states can be reconciled with those for a broad cross
section of countries if we allow for a notion of conditional conver-
gence in the underlying growth model. Some puzzles arise, however,
in reconciling the data with open-economy extensions of the model.
In particular, the rates of convergence found for income and product
across the U.S. states are similar, whereas theoretical reasoning sug-
gests some important differences.

Convergence in the Neoclassical Growth Model

In neoclassical growth models for closed economies, as presented by
Ramsey (1928), Solow (1956), Cass (1965), and Koopmans (1965), the
per capita growth rate tends to be inversely related to the starting
level of output or income per person. In particular, if economies are
similar in respect to preferences and technology, then poor econo-
mies grow faster than rich ones. Thus there is a force that promotes
convergence in levels of per capita product and income. Since the
model is familiar, we provide only a brief sketch.
The production function in intensive form is

9 f(k),11~~~~~~~~~1 (1)
where y and k are output and capital per unit of effective labor, Lext,
L is labor (and population), and x is the rate of exogenous, labor-
augmenting technological progress. (We assume the usual curvature
properties for the production function.) In a closed economy, k
evolves as
k =f(k) - c^-(8 + x + n)k, (2)
where c^= C/Lext, 8 is the rate of depreciation, and n is the growth
rate of L. The representative, infinite-horizon household maximizes
utility,
U = 1 u(c) ente-Ptdt, (3)

where c = CIL, p is the rate of time preference, and

uc C - 1
UW = 1- ' (4)
CONVERGENCE 225

with 0 > 0, so that marginal utility, u'(c), has the constant elasticity
-0 with respect to c. (We assume p > n + [1 - 0]x below to satisfy
the transversality condition.)
The first-order condition for maximizing U in equation (3) entails

l[f'(k) - 8 -p]. (5)

In the steady state, the effective quantities, 9, k, and c',do not change
and the per capita quantities, y, k, and c, grow at the rate x. The level
of k in the steady state satisfies

f'(k*) = 8 + p +Ox. (6)

If the economy starts with k below k*, then the usual analysis shows
that k monotonically approaches k* (see, e.g., Blanchard and Fischer
1989, chap. 2). We have shown (Barro and Sala-i-Martin 1991b, chap.
1) that the growth rate of capital per worker, k/k, declines monotoni-
cally toward the steady-state value, x. This property carries over un-
ambiguously to the growth rate of output per worker, 5/y, if the pro-
duction function is Cobb-Douglas, that is, if

f(k) = Aka,
= (7)
where 0 < a < 1. Thus if two economies have the same parameters
of preferences and technology, then the key result is that the initially
poorer economy-with a lower starting value of k-tends to grow
faster in per capita terms.
The transitional dynamics can be quantified by using a log linear-
ization of equations (2) and (5) around the steady state. The solution
for log[9(t)] in the log-linearized approximation to the model with a
Cobb-Douglas technology is

log[9(t)] = log[ 9(0)] *e-t + log(^*) *(1 -

COs (8)
where the positive parameter 13,which governs the speed of adjust-
ment to the steady state, is given by the formula

21 = 2 + 4(1 x)(P + 8 + Ox)

P + 8 Ox
+
~ 1/2 (9)
X [+ +OX(n + + x)] a

where j, p - n - (1 - O)x > 0.

The average growth rate of y over the interval between dates 0 and
T is
1~ SC ~~~~() I
*log [ ] = x + 1-e *log [Y ] (10)
T _Y(0) ~T J9(0)
226 JOURNAL OF POLITICAL ECONOMY

The higher A, the greater the responsiveness of the average growth

rate to the gap between log('*) and log[ y(0)], that is, the more rapid
the convergence to the steady state. The model implies conditional
convergencein that, for given x and 9*, the growth rate is higher the
lower y(0). The convergence is conditional in that y(0) enters in rela-
tion to 9* and x, which may differ across economies. In cross-country
regressions, it is crucial, but difficult, to hold fixed the variations in
9* and x in order to estimate P3.One advantage of the U.S. state
context is that the differences in 5* and x are likely to be minor, so
that conditional and absolute convergence need not be distinguished.
Because the crucial element for convergence in the neoclassical
model is diminishing returns to capital, the extent of these diminish-
ing returns-that is, the size of the capital-share coefficient a(in equa-
tion (7)-has a strong effect on P3.To assess the relation quantitatively
we use a set of baseline values for the other parameters: p = .05 per
year, 8 = .05 per year, n = .02 per year, x = .02 per year, and 0 =
1 (log utility). The value n = .02 per year is the average of population
growth for the United States over the long history. The other baseline
parameters come from estimates reported in Jorgenson and Yun
(1986, 1990). If we assume a( = .35-a capital share appropriate to
a narrow concept of physical capital (see, e.g., Maddison 1987)-then
equation (9) implies P3= .126 per year, which corresponds to a half-
life for the log of output per effective worker of 5.5 years. For ax =
.80, which might apply if capital is interpreted broadly to include
human capital, the value P3= .026 per year implies a half-life of 27
years. As at approaches unity, diminishing returns to capital disap-
pear, P3tends to zero, and the half-life tends to infinity.
The effects of the other parameters have been explored by Cham-
ley (1981) and King and Rebelo (1989).' Quantitatively, the most
important effect is that a lower 0 (increased willingness to substitute
intertemporally) raises P3.Another result is that the parameter A in
equation (7) does not affect P3.Thus the convergence coefficient P3
can be similar across economies that differ greatly in levels of per
capita product because of differences in the available technique (or
in government policies or natural resources that amount to differ-
ences in the parameter A).
The main result for the subsequent analysis is that the baseline
specification-including at = .35-generates a short half-life and a
rapid speed of adjustment. The speeds of adjustment that we estimate
empirically are much slower: , is in the neighborhood of .02 per
year. The theory conforms to the empirical findings only if we assume
parameter values that depart substantially from the baseline case.
1
Sato (1966) presents a related analysis for a model with a constant saving rate.
CONVERGENCE 227

One possibility is a value of axaround .8, that is, in the range in which
the broad nature of capital implies that diminishing returns set in
slowly. We can reduce the required value of catto around .5 if we
assume very high values of 0 (in excess of 10) and a value of 8 close
to zero.

Setup of the Empirical Analysis

Consider a version of equation (10) that applies for discrete periods
to economy i and is augmented to include a random disturbance:

log -te-N)
aYz(I log(y t )- (t- 1)] + i (11)

where = + (1 - e-O)log(fin, and ui is a disturbance term.

Althougha. thex. coefficient ,3 can vary across economies, we neglect
these differences in our analysis. This assumption is tenable for the
U.S. states, which are likely to be similar in terms of the underlying
parameters of technology and preferences. Also, as mentioned be-
fore, the theory implies that pure differences in the level of technol-
ogy do not affect P3.Thus 1Pcan be similar for economies that are
very different in other respects.
In the application to the U.S. states, we assume that the coefficient
a, in equation (11) is the same for all i; that is, we assume that the
steady-state value, fi* and the rate of technological progress, xi, do
not differ across states. The time trend, x, . (t - 1), is then also the
same for all i. The conditions a, = a and x, = x in equation (11) imply
that poor economies tend to grow unconditionally faster than rich
ones if P > 0. Because the coefficient on log(y,,t-1) is 1 - em, which
is between zero and one, the convergence is not strong enough to
eliminate the positive serial correlation in log(yit). Put alternatively,
in the absence of random shocks, convergence to the steady state is
direct and involves no oscillations. This property reflects the absence
of overshooting in the neoclassical growth model.
Convergence in the sense that poor economies tend to grow faster
than rich ones, which corresponds to P > 0 if ai and xi are the same
for all i in equation (11), does not necessarily imply that the cross-
economy dispersion of log(yit) declines over time. The effect from P3
> 0, which tends to reduce dispersion, is offset by random shocks,
uit, which tend to raise dispersion. If uit has zero mean and variance
ar and is distributed independently over time and across economies,
then the cross-economy variance of log(yit), denoted a 2, evolves as
a2 = (e -20)o-2_ 1 + o32 (12)
 228 JOURNAL OF POLITICAL ECONOMY

which implies

2 = U
art 1 _ +
Oro2 1e-e2 -
____
1 (13)

(We assume here that the cross section is large enough so that the
sample variance of log[yit] corresponds to the population variance,
(rt2.) Equation (13) implies that (y2 monotonically approaches the
steady-state value, or2 = a2/( - e -2*), which rises with o2 but declines
with P3.The variance orQ2 falls (or rises) over time if the initial value
a 2 is greater than (or less than) or2. Thus a positive coefficient P3does
not ensure a falling or2.
Shocks that have common influences on subgroups of countries or
regions, such as-harvest failures and oil shocks, imply that ud in equa-
tion (11) would'not be independent of ujt for j 1i. An important
example of this kind of shock from U.S. history is the Civil War,
which had a strong adverse effect on the southern states relative to
the northern states. We can handle this type of situation by writing
the error term, uit, in equation (11) as the sum of an aggregate influ-
ence and an independent disturbance:

(
log )= a- (1 - e ) *[log(yit-,)
) Y~~~~~,t
-x* (t - 1)] + 4dist+ Vit,

where st is an aggregate shock, which has zero mean and variance

2, and 4Xmeasures the effect of the aggregate disturbance on the
growth rate of economy i. We assume that, with 4ist held constant,
the error term, vit, is cross-sectionally and serially independent with
zero mean and constant variance or2.2
We assume that the coefficients Xi in equation (14) have mean +
and variance a 2 and are distributed independently of vit. If log(yi t- 1)
and Xi are uncorrelated, then estimates of the coefficient P3in equa-
tion (14) would not be systematically related to the realization of st
because the composite error term, uit = 4ist + vit, is uncorrelated
with the regressor, log(y' ,t- 1). Suppose, alternatively, that
cov[log(yit- ), Xi] > 0; for example, if a positive st represents an
increase in the relative price of oil, then economies that produce a
lot of oil (4h > 4) tend to have high values of yj toll In this case, the
least-squares estimate of the coefficient on log(yi t- 1) in equation (14)

2 The specification in eq. (14) means that realizations of st effectively shift crain eqq.
(12) and (13). Thus the approach of 4r2to a steady-state value need no longer be
monotonic. We plan in future research to analyze the time series of cr2 for the U.S.
states.
CONVERGENCE 229

is biased for a given realization of s,. For example, if oil-producing

economies have relatively high values of yi,- 1, then least-squares pro-
cedures tend to underestimate 13for a period in which the oil price
rises.3
In the empirical analysis, we include variables that we think hold
constant the effects of aggregate shocks, st, on economy i's growth
rate. One reason to add these variables is to achieve cross-sectional
independence of the error terms, vi, in equation (14): the composite
error, ud = 4ist + vit, would not exhibit this independence. The
second purpose is to obtain consistent estimates of the coefficient ,
conditional on the realizations of st.

The Data for the U.S. States

We have two measures of per capita income or product across the
U.S. states. The first is per capita personal income. The U.S. Com-
merce Department has published annual data on nominal personal
income for the 48 continental states since 1929 (see Bureau of Eco-
nomic Analysis [1986] and recent issues of Survey of CurrentBusiness).
We use the figures that exclude transfer payments from all levels
of government. Easterlin (1960a, 1960b) provides estimates of state
personal income for 1840 (29 states or territories), 1880 (47 states or
territories), 1900 (48 states or territories), and 1920 (48 states). These
data also exclude transfer payments.
We lack useful measures of price levels or price indexes for individ-
ual states. Therefore, we deflate the nominal values for each state by
the national index for consumer prices. Since we use the same price
deflator for each state in a single year, the particular deflator that we
use affects only the constant terms in the subsequent regressions. The
use of the same deflator for each state introduces two types of poten-
tial measurement error. First, if relative purchasing power parity does
not hold across the states, then the growth rates of real per capita
income are mismeasured. Second, if absolute purchasing power par-
ity does not hold, then the levels of real per capita income are mis-
measured.
The second type of data is per capita gross state product (GSP),
which is available annually for each state from 1963 to 1986 (see
Renshaw, Trott, and Friedenberg 1988). This variable, which is anal-

3 We assume here that yi represents either real per capita income for residents of
economy i (corresponding to the data on state personal income) or the real per capita
income derived from production of goods and services in economy i (corresponding
to the figures on gross state product). Hence, changes in relative prices show up directly
as changes in yi,; e.g., if no quantities change, then an increase in the relative price of
oil generates a high growth rate of yt, for economies that produce a lot of oil.
230 JOURNAL OF POLITICAL ECONOMY

ogous to gross domestic product (GDP), measures factor incomes de-

rived from production within a state. We deflate the nominal figures
by the aggregate GSP deflator for the year. (This deflator is close to
that for U.S. GDP.) Since we use a common deflator for each state at
a point in time, the particular deflator chosen is again of no conse-
quence. We should stress, however, that the GSP figures that we use
are not quantity indexes, but rather represent the incomes accruing
to factors from the goods and services produced within a state.
The main differences between state personal income and GSP in-
volve capital income. Personal income includes corporate net income
only when individuals receive payment as dividends, whereas GSP
includes corporate profits and depreciation. (Neither concept in-
cludes capital gains.) Most important, GSP attributes capital income
to the state in which the business activity occurs, whereas personal
income attributes it to the state of the asset holder.4

Evidence on Convergence for the U.S. States

We use the data on real per capita income or product, y, for a cross
section of the U.S. states, i = 1, .. ., N. Equations (10) and (11) imply
that the average growth rate over the interval between any two points
in time, to and to + T, is given by
1 Yi,to+?T e/i-ePT
-* logy ) = B - K T ) *iog(y09 ) + Ui'toto+T, (15)

where uit0t0+T is a distributed lag of the error terms, u-t, between

dates to and to + T.5 The constant term is B = x + [(1 - e IT)T]
[log('*) + xto], which is independent of i because we assumed f* -
9* and xi = x. The coefficient B shifts because of the trend in technol-
ogy with a change in the starting date, to.
The coefficient on log(yt,) in equation (15) is -(1 - e-T)IT,
which declines in magnitude with the length of the interval, T, for a
given P3.As T gets larger, the effect of the initial position on the
average growth rate gets smaller; as T tends to infinity, the coefficient
tends to zero. We estimate P3nonlinearly to take account of the associ-
ated value of T in the form of equation (15). Therefore, we should
obtain similar estimates of P3regardless of the length of the interval.
Table 1 contains nonlinear least-squares regressions in the form of

4 Some of these locational considerations apply also to labor income, although-

except for a few cities-the location of a business and the residence of the workers
are typically in the same state.
5 The error term is lIT times the sum for T between zero and T of the error
terms, Uito+T, weighted by e-(TT).
 TABLE 1
CROSS-STATE REGRESSIONS FOR PERSONAL INCOME

Sectoral
Composition
Sample (Sit) R v

1. 1880-1988 .0175 ... .92 .0014

(.0046)
2. 1880-1900 .0224 ... .62 .0054
(.0040)
3. 1900-1920 .0209 ... .67 .0062
(.0063)
4. 1920-30 -.0122 ... .43 .0111
(.0074)
5. 1930-40 .0127 ... .36 .0075
(.0051)
6. 1940-50 .0373 ... .86 .0057
(.0053)
7. 1950-60 .0202 ... .49 .0048
(.0052)
8. 1960-70 .0135 ... .68 .0037
(.0043)
9. 1970-80 .0119 ... .36 .0056
(.0069)
10. 1980-88 - .0005 ... .51 .0103
(.0114)
11. Nine periods, .0189 ... ... ...
,Brestricted* (.0019)
12. 1880-1900 .0268 - .0161 .65 .0053
(.0048) (.0079)
13. 1900-1920 .0269 - .0214 .71 .0060
(.0075) (.0094)
14. 1920-30 .0218 - .0936 .64 .0089
(.0112) (.0175)
15. 1930-40 .0141 2.43 .46 .0070
(.0048) (.81)
16. 1940-50 .0362 -.40 .87 .0057
(.0055) (.57)
17. 1950-60 .0313 .42 .65 .0041
(.0055) (.09)
18. 1960-70 .0194 .55 .71 .0036
(.0052) (.25)
19. 1970-80 .0139 .25 .36 .0056
(.0076) (.37)
20. 1980-88 .0196 1.35 .73 .0077
(.0106) (.22)
21. Nine periods, .0249 individual ...
,Brestricted* (.0021)
22. 1840-80 .0254 ... .91 .0030
(.0067)
NOTE.-Standard errors of coefficients are shown in parentheses. Regression 22 has 29 observations, regressions
1 and 2 have 47 observations (excluding Oklahoma), and regression 12 has 46 observations (excluding Oklahoma
and Wyoming). All others have 48 observations. The dependent variable is the growth rate of real per capita
personal income exclusive of transfers over the indicated sample period. Each regression includes a constant and
three regional dummy variables, south, midwest, and west. (Regression 22 includes only south and midwest.) The
coefficient d applies to log(yiO), where yi o is real per capita personal income at the start of the period. The sectoral
composition variable, Si, is described in the text. The regressions denoted nine periods, 3 restricted use nonlinear,
iterative weighted least squares, with the coefficient 3 constrained to be equal for all nine subperiods. Individual
coefficients are estimated for each subperiod for the constant, regional dummies, and the sectoral composition
variable.
* For line 11, the log likelihood ratio is 32.1 (p-value = .000); for line 21, it is 13.9 (p-value = .084). The
likelihood ratio statistic refers to the hypothesis of equality for the 3 coefficients. Under the null hypothesis, this
statistic is distributed as x2 with eight degrees of freedom.
232 JOURNAL OF POLITICAL ECONOMY

equation (15) for the U.S. states or territories and for various time
periods.6 Aside from log(yitO), each regression includes a constant
and three regional dummy variables: south, midwest, and west. (To
save space, the estimated coefficients for the constant and the regional
dummies are not shown in the table.) Because the regional dummies
are held constant, the effect of initial per capita income does not
reflect purely regional differences, such as the southern states' catch-
ing up with the northern states.7
For the longest interval, 1880-1988 (for 47 observations), the esti-
mated convergence coefficient shown in line 1 of table 1 is B= .0 175
(standard error = .0046). Figure 1 shows the dramatic inverse rela-
tion between the average growth rate from 1880 to 1988 and
log(y1880): the simple correlation is -.93.
The full time series for yi, (1880, 1900, 1920, and annually from
1929) potentially provides more information about the coefficient J3.
For a smaller value of T, however, the error term in equation (15),
Uitoto+T, represents an average of shocks over a shorter interval.
Therefore, the estimates become more sensitive to the specification
of the error process. In particular, if there is serial persistence in the
error term-, uit then the correlation between ui0toIto+T and log(yi to) is
likely to be negligible for large T but substantial for small T. For this
reason, we have not attempted to use the full annual time series that
starts in 1929.
Lines 2-10 of table 1 show estimates of P3for nine subperiods of
the overall sample: 1880-1900, 1900-1920, 10-year intervals from
1920 to 1980, and 1980-88. (There are 47 observations for the first
subperiod and 48 for the others.) Each regression includes a constant
and the three regional dummies. The results show values of 1 that
range from -.0122 (.0074) for 1920-30 to .0373 (.0053) for
1940-50.
If all nine subperiods are restricted to have a single value for ,
then the estimate is 13= .0 189 (.0019) in line 11. This estimation
allows each subperiod to have individual coefficients for the constant
and the regional dummies.8 The joint estimate of P3is close to the
value .0175 estimated for the single interval 1880-1988. But, as
would be expected, the standard error from the joint estimation,

6 See App. A for a discussion of the effects

of measurement error in Yiton the
estimates of P.
7The estimated 1 convergence across regions turns out to be similar to that within
regions (see Barro and Sala-i-Martin 1991a).
8 It would be possible to restrict the constants if it were maintained that each state
experienced exogenous technological progress at the constant rate x. We could then
use the whole sample to estimate a single constant and the value of x. We have not
imposed these restrictions because we have no reason to think that the rate of tech-
nological change would be the same over all time periods.
CONVERGENCE 233
0.025
FL VA

_
|~~~~~~~~A~
0.02

Ui
WV
~~~~~~~~~~~MD

8 0.015

C!,~~~~~~~~~~~~~~~~~~~~~~C
O 0.01

CL _AZ\

0.005--
.0.6 40.2 0.2 0.6 1 1.4 1.8

LOG(1880PERCAPITAPERSONALINCOME)
FIG. 1.-Growth rate from 1880 to 1988 vs. 1880 per capita income

.00 19, is a good deal smaller than that, .0046, found for the single
interval. The problem with the joint estimate is that the data reject
the hypothesis that the coefficient I3is the same for the nine subperi-
ods. The likelihood ratio statistic for this hypothesis, 32.1, is well
above the 5 percent critical value from the x2 distribution with eight
degrees of freedom of 15.5 (p-value = .000).
The unstable pattern of X coefficients across subperiods can reflect
aggregate disturbances that have differential effects on state incomes,
as represented by the term 4ist in equation (14). For example, during
the 1920s, the ratio of the wholesale price index for farm products
to the overall consumer price index fell at an average annual rate of
3.5 percent. The agricultural states also had below-average per capita
personal income in 1920: the correlation of log(y1920) with the share
of national income originating in agriculture in 1920 was - .67. Thus
the estimated coefficient, IB= -.0122, for the 1920-30 period in
table 1 likely reflects the tendency of the poorer states to be agricul-
tural and therefore to experience relatively low growth in this decade.
This effect reverses for the 1940-50 decade, when the ratio of the
wholesale price index for farm products to the overall consumer price
index grew at an average annual rate of 9,5 percent.
To hold constant this type of effect, we construct a variable that
measures the sectoral composition of income in each state. For the
subperiods that begin since 1930, we use a breakdown of the sources
of labor income (including income from self-employment) into nine
234 JOURNAL OF POLITICAL ECONOMY

categories: agriculture; mining; construction; manufacturing; trans-

portation and public utilities; wholesale and retail trade; finance, in-
surance, and real estate; services; and government and government
enterprises. For each subperiod, we construct a sectoral composition
variable for state i:
9 /
jt T), (16)
Sit wijt log(
j=1 j

where wit is the weight of sectorj in state i's personal income at time
t and Yjtis the national average of personal income that originates in
sector j at time t, expressed as a ratio to national population at time
t. Aside from the effect of changing sectoral weights within a state,
the variable sit would equal the growth rate of per capita personal
income in state i between years t and t + T if each of the state's
sectors grew at the national average rate for that sector. In particular,
the variable reflects shocks to agriculture, oil, and so forth in a way
that interacts with state i's concentration in the sectors that do rela-
tively well or badly in terms of income because of the shocks.
We think of the variable sit as a proxy for common effects related
to sectoral composition in the error term in equation (15). Note that
sit depends on contemporaneous realizations of national variables, but
only on lagged values of state variables. Because the impact of an
individual state on national aggregates is small, sit can be nearly exog-
enous with respect to the current individual error term for state i. In
any event, we assume that, with sit held constant, the error terms are
independent across states and over time.
For the subperiods that begin before 1930, we lack detailed data
on the sectoral composition of personal income, but we have data on
the fraction of national income originating in agriculture. For these
subperiods, we use this fraction as a measure of sit. Note that the
different methods of construction and the differing behavior of ag-
ricultural relative prices mean that the coefficients of the variable sit
will vary from one subperiod to another. Therefore, we estimate a
separate coefficient on sit for each subperiod.
Lines 12-20 in table 1 add the variable sit to the growth rate regres-
sions for each subperiod. (The first subperiod has 46 observations
and the others have 48.) As before, these regressions include
log(yito), a constant, and three regional dummies. Not surprisingly,
the estimated coefficients on the variable sit for the post-1930 sub-
periods are typically positive. That is, states in which income origi-
nates predominantly in sectors that do well at the national level tend
to have higher per capita growth rates. (The estimated coefficient for
the 1940-50 subperiod is negative, but not significantly so.) For the
CONVERGENCE 235

subperiods that begin before 1930, the negative estimated coefficient

on sit signifies that, with initial per capita income and region held
constant, agricultural states have lower per capita growth rates. This
pattern is especially clear for the agricultural price collapse in the
1920-30 decade: the estimated coefficient on sit is -.0936 (.0175).
For our purposes, the principal finding from the addition of the
sectoral composition variables is that the estimated P3coefficients be-
come much more stable across subperiods. The range is now .0139
(.0076) for 1970-80 to .0362 (.0055) for 1940-50. Line 21 shows
that the jointly estimated coefficient for the nine subperiods is .0249
(.0021). (This joint estimation allows each subperiod to have individ-
ual coefficients for sit as well as for the constant and the regional
dummies.) The likelihood ratio statistic for the equality of P coeffi-
cients across the nine subperiods is now 13.9, compared to the 5
percent critical value of 15.5. Thus if we hold constant the measures
of sectoral composition, we no longer reject the hypothesis of a single
f3 coefficient at the 5 percent level (p-value = .084).
The agriculture share variable, which was included to measure sit
for the earlier subperiods in table 1 (lines 12-14 and the joint esti-
mate in line 21), holds constant compositional effects on aggregate
state income that reflect shifts of persons out of agriculture and into
higher-productivity jobs in industry and services. If we add the agri-
culture share variable to the later subperiods, then the joint estimate
for nine subperiods becomes ,3 = .0224 (.0022), slightly less than the
value shown in line 21. This estimate of IBis virtually unchanged if
we include the change in the agriculture share over each subperiod
in the regressions. Thus convergence at a rate of about 2 percent per
year is net of effects from changes in agricultural shares.
In general, industry mix effects would matter for the results if
changes in income shares among sectors with different average levels
of productivity are correlated with initial levels of per capita income.
It is unclear that we would want to filter out all these effects to mea-
sure convergence, but, in any event, our examination of productivity
data from the post-World War II period indicates that shifts between
agriculture and nonagriculture would be the main effect of this type.
Since we already held constant the compositional effect for agricul-
ture, it is unlikely that industry mix effects are a major element in
the estimated convergence for state personal income.
The final result from table 1 is a regression with the 29 available
observations from 1840 to 1880.9 This regression includes a constant
9 Easterlin (1960a, p. 124 ff.) indicates that the data for 1840 do not cover income
originating in wholesale and retail trade; finance, insurance, and real estate; govern-
ment; and most other services. The figures that we use for 1880 in the 1840-80
regressions are comparable in coverage to those for 1840. This more limited coverage
236 JOURNAL OF POLITICAL ECONOMY

0.025

0.0215R

0.01 -

- C
0.01 R
CD~~~~~~~~~~DM MO ME
moos
0R

-0.005SOT

-1.3 -1.1 .0.9 40.7 .0.5 40.3 .0.1 0.1 0.3

LOG(1840PERCAPITAPERSONALINCOME)
FIG. 2.-Growth rate from 1840 to 1880 vs. 1840 per capita income

and two regional dummies (no western states are in the sample). We
exclude the variable s - because the data are unavailable. The estimate
in line 22 is 3 = .0254 (.0067), which accords with the estimate of
.0249 (.0021) for the subperiods that begin after 1880 (line 21).
Figure 2 plots the per capita growth rate from 1840 to 1880 against
log(yI840). A remarkable aspect of the plot is the separation of the
southern and nonsouthern states because of the Civil War. In 1840,
the southern and nonsouthern states differed little in terms of aver-
age per capita income: the unweighted average of 11 southern states
was 94 percent of that for 18 eastern and Midwestern states. But in
1880 a wide gap had appeared and the southern average was only
50 percent of the nonsouthern. The figure shows, however, that con-
vergence applies to the southern and nonsouthern states as separate
groups. That is, with the regional dummies held constant (which ef-
fectively hold constant the impact of the Civil War), there is a strong
negative correlation between the per capita growth rate and the initial
level of per capita income.
The Civil War affected states differentially, but, in contrast to the
shock to agriculture in the 1920s, the effect of the Civil War on state
per capita income had little correlation with the initial level of per

for 1880 comprises about half the income included in the measure that we used previ-
ously. In any event, the limited figures for 1840 are not comparable to the data for
years after 1880.
CONVERGENCE 237

capita income. For this reason, we do not get a very different point
estimate of I for the 1840-80 subperiod if we eliminate the regional
dummies: the estimate without these dummies is I = .0203 (.0126).
The fall in the R2 of the regression from .91 in line 22 of table 1
to .19 indicates, however, that the regional dummies have a lot of
explanatory power in this period!

Results with Gross State Product

Table 2 and figure 3 deal with the growth of per capita GSP for 48
states over the period 1963-86. Recall that the data are nominal GSP
divided by an aggregate, national price deflator. The growth rates

TABLE 2
CROSS-STATE REGRESSIONS FOR GROSS STATE PRODUCT

Sectoral
Sample Composition 2
Sample ,B(sit) R2 a
1. 1963-86 .0180 ... .48 .0038
(.0059)
2. 1963-69 .0154 ... .63 .0056
(.0060)
3. 1969-75 .0406 ... .41 .0120
(.0162)
4. 1975-81 - .0285 ... .17 .0139
(.0130)
5. 1981-86 .1130 ... .62 .0168
(.0244)
6. Four periods, .0211 ... ... ...
f restricted* (.0053)
7. 1963-69 .0157 .18 .63 .0056
(.0060) (.25)
8. 1969-75 .0297 1.56 .74 .0081
(.0101) (.20)
9. 1975-81 .0258 1.74 .78 .0072
(.0108) (.15)
10. 1981-86 .0238 1.73 .92 .0079
(.0091) (.13)
11. Four periods, .0216 individual ... ...
,3 restricted* (.0042)
12. 1963-86 .0222 .63 .54 .0036
(.0065) (.27)

NOTE.-All regressions have 48 observations. The dependent variable is the growth rate of real per capita GSP
(nominal GSP per capita divided by the national deflator for GSP). The regressions denoted four periods, P
restricted use nonlinear, iterative weighted least squares, with the coefficient J constrained to be equal for the four
subperiods. See also the notes to table 1.
* For line 6, the log likelihood ratio is 31.2 (p-value = .000); for line II, it is 1.7 (p-value = .637). Under the
null hypothesis of equal coefficients, the likelihood ratio statistic is distributed as x2 with three degrees of freedom.
238 JOURNAL OF POLITICAL ECONOMY

0.032

\ ~~~~~~GA
NH

(O 0.027

a) 0.024- ME TX I-A

iii 0.Q1 SCT - 9 j <s W

MN
-S
0.021 SDco NJ

0.018 F\ CA

I< 0.015 - ID WA

I.
cc
WV UT
0012-
IL _

0.009-
0. 09 l
I I I I I I

1.6 1.8 2 2.2 2.4

LOG(1963 PER CAPITAGSP)

FIG. 3.-Growth rate from 1963 to 1986 vs. 1963 per capita GSP

therefore pick up changes in relative prices that interact with a state's

composition of production. However, the structural variable, sit, holds
constant these effects from changes in relative prices.
For the full sample, 1963-86, the estimated convergence coeffi-
cient in line 1 of table 2 is 13= .0 180 (.0059). This regression includes
a constant and the three regional dummy variables, but no measures
of sectoral composition. The regressions over subperiods (1963-69,
1969-75, 1975-81, and 1981-86 in lines 2-5) show marked instabil-
ity in A: it ranges from -.0285 in 1975-81 to .1130 in 1981-86. The
joint estimate of 1 for the four subperiods (line 6) is .0211 (.0053),
but the hypothesis of equal coefficients is rejected (p-value = .000).
We again add a measure of sectoral composition, sit, analogous to
that defined in equation (16). The difference is that the data allow
us to disaggregate into 54 sectors for the origination of GSP. Lines
7-10 in table 2 show that the 1 coefficients are similar across the
subperiods when the variable sit is held constant. The joint estimate
in line 11 is P = .0216 (.0042), and the hypothesis of stability across
the four subperiods is accepted at the 5 percent level (p-value = .64).
Some of the instability in the 1 coefficients with the GSP data relates
to the movements in oil prices. Oil prices and, hence, the incomes of
oil states rose substantially during the subperiod 1975-81. Moreover,
the oil states were already relatively high in per capita GSP by 1975:
the correlation of per capita GSP with the share of GSP originating
in crude oil and natural gas was .4. The tendency of the rich oil states
to grow at relatively high rates upsets the usual convergence pattern
CONVERGENCE 239
and thereby leads to the negative value for A, -.0285, shown for
1975-81 in line 4 of table 2.'? But when sectoral composition is held
constant in line 9, the value of 13for 1975-81 is similar to that found
in the other periods.
For the 1981-86 period, the key elements are the sharp decline in
oil prices and the high correlation, .7, between per capita GSP and
the share of GSP originating in oil and natural gas in 1981. The
tendency for oil states to do relatively badly in 1981-86 leads to an
exaggerated convergence coefficient, 13= .1130, in line 5.11 Again,
the inclusion of the variable sit in line 9 leads to a normal value for

Barro and Sala-i-Martin (199la) disaggregate the nonagricultural

part of GSP into value added per worker for eight sectors. The main
finding is that convergence shows up significantly within these sectors
of production, especially for manufacturing. For the nonmanufactur-
ing sectors, the overall estimate of ,3 is somewhat less than 2 percent
per year, whereas for manufacturing the estimate is over 4 percent
per year. The main inference from these results is that poorer states
grow faster not only in terms of overall GSP per person, but also in
terms of labor productivity within various sectors of production.
Thus, as suggested before for personal income, the findings on con-
vergence cannot be explained by changes over time in the composi-
tion of production.

Income versus Product

In a closed-economy growth model, the convergence properties of
income and product must coincide. Perhaps surprisingly-because
the U.S. states do not look like closed economies-the empirical esti-
mates of f for personal income are nearly equal to those for GSP. If
the estimation for personal income is limited to a time span similar
to that covered by GSP-namely the three subperiods 1960-70,
1970-80, and 1980-88-then the joint estimate is ,3 = .0181 (.0040).
Although this point estimate is less than that, .0216 (.0042), shown
for GSP in table 2, line 11, the principal finding is that the estimates
are close.
The assumptions of a closed economy are implausible for the U.S.

10This argument does not apply to the subperiod 1969-75 (line 3 of table 2).
Although the oil price rose substantially over this period, the oil states did not
have especially high values of per capita GSP in 1969.
"' The results for personal income over the period 1980-88 (table 1, line 10) do not
show the same pattern. The main difference is that the correlation in 1980 of the
logarithm of per capita personal income with the share of income originating in oil
and natural gas is close to zero.
240 JOURNAL OF POLITICAL ECONOMY

states: goods and technologies flow across borders, residents of one

state can borrow from residents of other states, and internal migra-
tion is possible. In Barro and Sala-i-Martin (1991b, chap. 2), we ex-
tend the neoclassical growth model to allow for internationally trad-
able goods and a global capital market. These features create a sharp
distinction between domestic product and income or, equivalently,
between domestic capital stock and assets. If technologies are the
same, then an economy's per capita capital stock and output converge
rapidly to those prevailing in other economies. In contrast, even if all
economies have the same parameters of preferences and technolo-
gies, per capita incomes do not converge because each small economy
faces constant returns on the global capital market. Thus our empiri-
cal findings-that rates of convergence are similar for income and
product across the U.S. states-are puzzling from the perspective of
this theory. We offer here some conjectures that may help to resolve
this puzzle.
We have modified the analysis along the lines of Cohen and Sachs
(1986) to allow for a ceiling on the ratio of an economy's external debt
to its capital stock. This restriction on credit markets is reasonable if
the capital stock represents the collateral that secures the debt. If we
interpret capital broadly to include human capital, then this frame-
work applies to the U.S. states if the residents or government of a
state cannot borrow nationally to finance all their desired expendi-
tures on education or other forms of investment in human capital.
The key result from the addition of the borrowing constraint is that
domestic product behaves eventually like national income. Hence, the
convergence properties of product and income can be similar, as in
our empirical results.
If technologies (i.e., anything represented by the coefficient A in
eq. [7]) differ across economies, then mobility of capital can create
divergence of per capita output and capital stocks. Economies with
higher k tend to have higher values of A, and the higher A offsets
the effect of diminishing returns in the determination of capital's
marginal product. Therefore, capital (physical or human) may move
from poorer to richer economies, and it is no longer clear theoreti-
cally that the convergence coefficient for product would exceed that
for income. Once we allow for differences in technologies, we also
have to consider the diffusion of technology across economies, along
the lines of Nelson and Phelps (1966). The potential to imitate is
another reason for poor, follower economies to grow at relatively
high rates.
We have extended the neoclassical growth model to allow for mi-
gration of persons, another force that promotes convergence of per
capita product and income across economies. Sala-i-Martin (1990,
CONVERGENCE 241

table 5.2) and Barro and Sala-i-Martin (199la) relate net migration
for the U.S. states to initial values of per capita personal income over
subperiods of the interval from 1900 to 1987. These studies confirm
that net in-migration is positively related to initial per capita income.
But the results also show that the estimated convergence coefficients,
A, are little affected by the inclusion of net migration as an explana-
tory variable in the growth rate equations. Moreover, we have shown
that the minor interplay between migration and convergence is quan-
titatively consistent with the neoclassical growth model (extended to
allow for migration), given the estimated sensitivity of migration to
income differentials.
We leave as an unresolved puzzle the similar estimates for the rates
of convergence of per capita income and product. We think that a
resolution of this puzzle will involve the construction of an open-
economy growth model that satisfactorily incorporates credit mar-
kets, factor mobility, and technological diffusion.

Comparisons with Findings across Countries

In this section we compare our findings for the U.S. states with analo-
gous results across countries. It is well known that growth rates of
real per capita GDP are uncorrelated with the starting level of real
per capita GDP across a large group of countries in the post-World
War II period. Barro (1991) uses the Summers-Heston (1988) data
set along with other data to analyze the growth experiences of 98
countries from 1960 to 1985. (The limitation to 98 countries rather
than the 114 market economies with Summers-Heston GDP data
from 1960 to 1985 comes from the lack of information on variables
other than GDP.) Line 1 of table 3 shows that a regression for the 98
countries in the form of equation (15) leads to the estimate 1 =
-.0037 (.0018). The dependent variable is the growth rate of real
per capita GDP from 1960 to 1985. The only independent variables
are a constant and the log of 1960 per capita GDP, log(y1960). The
main finding, also depicted in figure 4, is the lack of a close relation-
ship between the growth rate and log(y1960). In fact, the convergence
coefficient 1 has the wrong sign; that is, there is a small tendency for
the initially rich countries to grow faster than the poor ones after
1960.
These cross-country results contrast sharply with the findings dis-
cussed earlier for the U.S. states. Figures 1 and 3 and tables 1 and 2
showed that, particularly over the longer samples, there is a clear and
substantial negative correlation between starting per capita income
or product and the subsequent growth rate. Line 5 of table 3 uses
a specification for the U.S. states that parallels the one used for
 TABLE 3
COMPARISON OF REGRESSIONS ACROSS COUNTRIES AND U.S. STATES

Additional
Sample Variables R2 6
1. 98 countries, -.0037 no .04 .0183
1960-85 (.0018)
2. 98 countries, .0184 yes .52 .0133
1960-85 (.0045)
3. 20 OECD countries, .0095 no .45 .0051
1960-85 (.0028)
4. 20 OECD countries, .0203 yes .69 .0046
1960-85 (.0068)
5. 48 U.S. states, .0218 no .38 .0040
1963-86 (.0053)
6. 48 U.S. states, .0236 yes .61 .0033
1963-86 (.0013)

NOTE.-The dependent variable in regressions 1-4 is the growth rate of real per capita GDP from 1960 to 1985;
in regressions 5 and 6 it is the growth rate of real per capita GSP (the variable used in table 2) from 1963 to 1986.
The coefficient P applies in regressions 1-4 to the logarithm of real per capita GDP in 1960, and in regressions 5
and 6 to the logarithm of real per capita GSP in 1963. Each regression also includes a constant. The additional
variables included in regressions 2 and 4 are the primary and secondary school enrollment rates in 1960, the
average ratio of government consumption expenditure (standard figures less spending on defense and education)
to GDP from 1970 to 1985, the average number of revolutions and coups per year from 1960 to 1985, the average
number of political assassinations per capita per year from 1960 to 1985, and the average deviation from unity of
the Summers-Heston (1988) purchasing power parity ratio for investment in 1960. See Barro (1991) for details on
these variables. The additional explanatory variables included in regression 6 are regional dummies, the sectoral
composition variable, si, and the fraction of workers in 1960 that had accumulated some amount of college educa-
tion. The 20 OECD countries (the original membership in 1960) are Austria, Belgium, Canada, Denmark, France,
Germany, Greece, Iceland, Ireland, Italy, Luxembourg, Netherlands, Norway, Portugal, Spain, Sweden, Switzer-
land, Turkey, United Kingdom, and United States.

0.08 -
44

0.07 -
LO 33
OD
0.06 - 36 37

6CD
0) 0.05 -2
40 55 50
- 0.04 - 23 63 64
6 26 W85 8 42
36
I 0.03 - 3 4163542
5 1
90 67 ~5170
O 0.0~~~~~~~2 - 5172195 ,,,
87

0
(9
0.02 - 24
274
3
34
7 97 7#459 9~~~~~~~~~~~~~
8H32

< 0.01 -
13 79
(!) 8196 a
<
0~~~~~~~~~~~58 25
73 93
C,
-0.0 - l1l9
W -0.021
-
2.0 - 131
-2 02

LOG(1960 PER CAPITAGDP)

FIG. 4.-Growth rate from 1960 to 1985 vs. 1960 per capita GDP, sample of 98
countries (listed in App. B).
CONVERGENCE 243
the countries. The variables are based on GSP over the time pe-
riod 1963-86, and the regression includes only log(y1963) and a con-
stant as regressors. The estimate in this case is ,3 = .0218 (.0053).
Barro (1991, table 1, fig. 2) shows that a significantly negative par-
tial relation between the per capita growth rate from 1960 to 1985
and log(y1960) emerges for the 98 countries if some other variables are
held constant. The set of other variables in the main results consists of
primary and secondary school enrollment rates in 1960, the average
ratio of government consumption expenditure (exclusive of defense
and education) to GDP from 1970 to 1985, proxies for political stabil-
ity, and a measure of market distortions based on purchasing power
parity ratios for investment goods. If we include these variables for
the 98 countries in the form of equation (15), then line 2 of table 3
shows that the estimated convergence coefficient becomes ,B = .0184
(.0045). This estimate of 13is no longer very much below the cross-
state value shown in line 5 of the table.
The theoretical relation in equation (15) predicts conditional con-
vergence, that is, a negative relation between log(yito) and the subse-
quent growth rate if we hold constant the steady-state position,
log(j%>, and the steady-state growth rate, xi. (The constant B in
eq. [15] depends on log[y1 and xi.) The theory implies that the rela-
tion between log(yito) and the growth rate will be negative unless the
correlation between log(yiO) and the two omitted factors, log(9,)
and xi, is substantially positive.
The U.S. states are likely to be reasonably homogeneous with re-
spect to the steady-state values log(9i) and xi. That is, the differences
in initial positions, log(yiO), may be relatively much greater. (This
condition is especially compelling if the initial differences reflect ex-
ogenous events, such as wars, world agricultural harvests, and oil
shocks.) In this case, the negative relation between the growth rate
and log(y 00)would show up even if the differences in the steady-state
values are not held constant: conditional and absolute convergence
would coincide. The result for 13shown in line 5 of table 3 is consistent
with this perspective.
In contrast, the sample of 98 countries likely features large differ-
ences in the steady-state values, log(j1) and xi, that is, in the underly-
ing parameters of technology and preferences (and natural resources
and government policies) that determine these long-run values. The
absence of substantial labor mobility across countries reinforces the
possibility of substantial divergences in these steady-state values. The
correlation of log(y ,,O)with log(9 is likely to be substantially positive;
that is, economies with higher steady-state values of output per effec-
tive worker would have followed a path that led them today to higher
levels of output per person. Similarly, the correlation of log(yito) with
244 JOURNAL OF POLITICAL ECONOMY

xi is likely to be positive. Because of these relations, the simple correla-

tion between the growth rate and log(yi to) could be close to zero, as
indicated by the data in figure 4. (This point is made by King and
Rebelo [1989, pp. 12-13].) On the other hand, if we include addi-
tional variables that hold constant some of the cross-country varia-
tions in log(yi) and xi, then the partial relation between growth rate
and log(yito) should become more negative. We interpret the addi-
tional variables that we added to the cross-country regression (table
3, line 2) in this manner. Accordingly, we view the estimate of 0 in
this regression-which is no longer very much below the values from
the cross-state regressions-as coming closer to the theoretical con-
vergence coefficient.
We can evaluate these arguments further by considering a group
of relatively homogeneous countries, the 20 original members of the
OECD.12 Figure 5 shows that the per capita growth rate is negatively
related to the log of initial per capita GDP for this group of countries.
The regression in line 3 of table 3 includes only a constant and
the log of 1960 per capita GDP. The estimated convergence coeffi-
cient is 13= .0095 (.0028), which is significant and has the expected
sign. The magnitude is, however, about half that applicable to the
U.S. states (line 5). Our interpretation is that the OECD countries are
intermediate between the 98-country group and the U.S. states in
terms of the extent of cross-country variation in steady-state values,
log(j^5 and xi, relative to the variation in initial positions, log(yito).
Line 4 of the table shows that the estimate for the OECD countries
becomes 1 = .0203 (.0068) when the additional variables discussed
before are added to the regression. This estimate does not differ
greatly from the comparable value for 98 countries, .0184 (.0045) in
line 2.
We have also explored in a preliminary way the addition of vari-
ables as proxies for the steady-state values, log(5^5:and xi, in the cross-
state regressions. One variable that has a significantly positive influ-
ence on the growth rate is the fraction of the work force in 1960 that
had accumulated some amount of college education.'3 We added this
variable along with the regional dummies and sectoral composition
variable, sit, that we discussed before. Line 6 of table 3 shows that
the estimated convergence coefficient becomes i = .0236 (.0013),
compared with .0218 (.0053) in line 5. Thus the inclusion of these
12 We exclude the four countries added after 1960 (Australia, Finland, Japan, and
New Zealand) because of the possibility that the extension of membership was endoge-
nous and related to the growth experience.
13 The data on educational attainment come from various issues of StatisticalAbstract.

We have not had much success in finding growth rate effects related to cross-state
differences in government expenditures. Also, educational differences aside from col-
lege attainment were not important.
CONVERGENCE 245
0.0 -6

z 0.02 -

57QO QE 0. 1 1. 48
eo
Q1- 02

0..3 mom1 - , I, , , ,

LOG(1960 PER CAPITAGDP)

FIG. 5.-Growth rate from 1960 to 1985 vs. 1960 per capital GDP, OECD countries
(listed in App. B).

other variables has a positive, but minor, effect on the estimate of 1

across the U.S. states.
Overall, the impact of the additional variables on 13is greatest for
the 98 countries (.0184 in line 2 vs. - .0037 in line 1), next most
important for the 20 OECD countries (.0203 in line 4 vs. .0095 in
line 3), and least important for the 48 U.S. states. These findings are
consistent with the idea that, first, the other variables help to hold
constant cross-sectional differences in the long-run values,log(gas
and xi, and, second, that the ranking of the extent of these differences
(relative to the differences in log[yuno])goes from the 98 countries to
the 20 OECD countries to the 48 U.S. states.

Conclusions
Our empirical results document the existence of convergence in the
sense that economies tend to grow faster in per capital terms when
they are further below the steady-state position. This phenomenon
shows up clearly for the U.S. states over various periods from 1840
to 1988. Over long samples, poor states tend to grow faster in per
capita terms than rich states even if we do not hold constant any
variables other than initial per capita income or product. If we hold
constant the region and measures of sectoral composition, then the
speed of convergence appears to be roughly the same-around 2
percent per year-regardless of the time period or whether we con-
sider personal income or GSP.
246 JOURNAL OF POLITICAL ECONOMY

We find evidence of convergence for a sample of 98 countries from

1960 to 1985 only in a conditional sense, that is, only if we hold
constant variables such as initial school enrollment rates and the ratio
of government consumption to GDP. We interpret these variables as
proxies for the steady-state value of output per effective worker and
the rate of technological progress. If we hold constant these addi-
tional variables, then the estimated rates of convergence are only
slightly smaller than those found for the U.S. states.
The standard neoclassical growth model with exogenous techno-
logical progress and a closed economy predicts convergence. To
match our quantitative estimates, however, we have to assume under-
lying parameters for preferences and technology that depart substan-
tially from usual benchmark cases. In particular, for reasonable val-
ues of the other parameters, the model requires a capital share
coefficient, (x, in the neighborhood of .8. Lower values of a, which
imply that diminishing returns to capital set in more quickly, imply
-a more rapid rate of convergence than that revealed by the data.
If technologies are the same, then the introduction of a global
capital market tends to speed up the convergence for output but to
slow down the convergence for income. The empirical results for the
U.S. states indicate that the speed of convergence for output is only
slightly faster than that for income. At this point, we can reconcile
this finding with the theory only if we include elements of capital
market imperfections, such as a limited ability to borrow to finance
accumulations of human capital. Other elements of an open
economy-the mobility of labor and technology-tend to speed up
the predicted rate of convergence. Therefore, we require even higher
values of the capital share parameter, (x, to match the empirical re-
sults.
Some recent models of endogenous economic growth, such as Re-
belo (1991), assume constant returns to a broad concept of capital
that includes human capital. This specification corresponds to (x =
1.0 in the neoclassical model. As mentioned, our empirical results
indicate that the neoclassical model requires a value of (x of about .8
to fit the observed speeds of convergence. The difference between ax
= .8, where diminishing returns to capital set in slowly, and ax = 1.0,
where diminishing returns do not set in at all, may seem to be minor.
But the difference amounts to a half-life of 27 years in the former
case versus infinity in the latter. To put it another way, the conver-
gence coefficient 13= 2 percent per year, corresponding to ax - .8,
implies that the poor countries of sub-Saharan Africa should have
experienced growth of real per capita GDP from 1960 to 1985 at an
average rate above 6 percent per year, compared to 2 percent per
year for the United States, if the African countrieswere approachingthe
CONVERGENCE 247

same steady-statepath as that for the United States. (The actual average
growth rate of 0.8 percent per year for the sub-Saharan African coun-
tries is "explained" in the regression in line 2 of table 3 by the addi-
tional variables that proxy for steady-state positions.) The main point
is that a value for (x of .8 is very far from 1.0 in an economic sense.
In open-economy versions of the neoclassical growth model, it is
possible to find convergence effects associated with technological dif-
fusion even if the returns to capital are constant (ot = 1). Also, in
closed-economy models with constant returns to a broad concept of
capital, convergence effects can reflect the working out of initial im-
balances among the various kinds of capital. For example, Mulligan
and Sala-i-Martin (1991) show that the per capita growth rate is in-
versely related to initial physical capital per worker for a given initial
quantity of human capital per worker. Thus we would like to break
down the observed convergence into various components: first, ef-
fects related to diminishing returns to capital and to imbalances
among types of capital in the context of a closed economy; second,
effects involving the mobility of capital and labor across economies;
and third, effects that involve the gradual spread of technology. The
present empirical results, which exploit only cross-sectional differ-
ences in growth rates, do not allow us to separate the observed con-
vergence patterns into these components. We hope to make these
distinctions in future research, which will also exploit the time-series
variations of growth rates.

Appendix A
Some Effects of MeasurementError
The regressionsshown in tables 1 and 2 can exaggeratethe estimatedconver-
gence coefficient, I, if real income or product is measured with error. Aside
from the usual measurement problems, one reason to expect errors is that
we divide all nominal values in each year by a common price index.
Equation (15) can be rewritten as

1T*(Yito+T)= B + T * log(yt ) + Uitt+T (A1)

Assume that the observed value at date t, log(yt), differs from the true value,
log(yit),by a random measurement error:
log(yld) = log(yzt) + nit, (A2)
For purely temporary measurement error, nit would be white noise. Then,
as is well known, the measurementerror in log(yi,o) implied by equation (A2)
leads to a bias toward zero in least-squaresestimation of the coefficient,
e-T/T, in equation (Al). Because the term e TIT in equation (Al) is decreas-
ing in I, the nonlinear estimate I provides a corresponding overestimateof
P in large samples.
248 JOURNAL OF POLITICAL ECONOMY

We can obtain a bound for the inconsistencyinduced by temporary mea-

surementerror. Equation(11) impliesthat the growthrate of income between
any two future dates, to + T and to + T, is given by
1 Yizto+T\ e- e-PT
T- l-?y T) T log(Y,7to) + Uito+,to+T, (A3)

where T > T > 0 and u- t +T to+T depends on the error terms, uit, between
dates to + T and to + T. Equation (A3) relates the growth rate from to + T
to to + T to the level of per capita income or product at an earlier time, to.
Note that equation (15) is the special case in which T = 0.
We assume that the measurement error, biqteis independent of lit,+t
for t 2? T. This condition holds for all T > 0 if is white noise but also applies
s

for large enough T to measurement error with some persistence over time.
We assume that hi tp is independent of uito+Tto+T. In this case, least-squares
estimation of equation (A3) leads to an underestimateof the magnitude of
the coefficient, (ed-t - e-T)/(T - T). We can show that this term is increasing
in I if I < [log(T/T)]/(T- T). In practice, we use the values T = 10 years
and T = 20 years or T = 5 years and T = 10 years. For the first pair of
values, the term (edit - e-T)/(T - T) is increasingin I if I < .07 per year;
for the second pair, the term is increasingin f if I < .14 per year. Therefore,
for these ranges of I and in large samples, the underestimateof the coeffi-
cient on log(yito) in equation (A3) corresponds to a large-sampleunderesti-
mate of P3.Because this bias is opposite in directionto that found for equation
(15), we can use regressions in the form of equation (A3) to bound the size
of the bias.
Consider the regressions for personal income in which each subperiod has
individual coefficients for the constant, three regional dummies, and the
sectoralcomposition variable,sit If we use only the five equal-lengthsubperi-
ods from 1930-40 to 1970-80, then the joint estimate P in the form of
equation (15) is .0244 (.0025), which is close to the value for nine subperiods
from 1880 to 1988 shown in line 20 of table 1. The comparableresult in the
form of equation (A3) with T = 10 years and T = 20 years is I = .0278
(.0049). Although we expected the asymptotic bias induced by temporary
measurement error to be positive in the first case and negative in the second,
the result for f turns out to be higher in the second case. (The theoretical
result can be affected by the inclusion of additional explanatoryvariablesin
the regressions.) In any event, we infer from the similarityof the two esti-
mates of P that temporary measurement error is unlikely to have a major
influence on the results.
For GSP, we use the three equal-length subperiods 1970-75, 1975-80,
and 1980-85. The joint estimate I in the form of equation (15) is .0280
(.0058), somewhathigher than that, .0216 (.0042), shown for four subperiods
from 1963 to 1986 in line 11 of table 2. With T = 5 years and T = 10 years,
joint estimation of equation (A3) over the three subperiods from 1970 to
1985 leads to the estimate I = .0366 (.0091). Again, in contrast to expecta-
tions, the estimated value in the second case exceeds that in the first case. But
the main inference is that the results are similarand, hence, that temporary
measurement error is unlikely to be important.
CONVERGENCE 249
Appendix B
Key for Countries in Figures 4 and 5
1. Algeria 51. Denmark
2. Botswana 52. Finland
3. Burundi 53. France
4. Cameroon 54. Germany
5. Central African Republic 55. Greece
6. Egypt 56. Iceland
7. Ethiopia 57. Ireland
8. Gabon 58. Italy
9. Ghana 59. Luxembourg
10. Ivory Coast 60. Malta
11. Kenya 61. Netherlands
12. Liberia 62. Norway
13. Madagascar 63. Portugal
14. Malawi 64. Spain
15. Mauritius 65. Sweden
16. Morocco 66. Switzerland
17. Nigeria 67. Turkey
18. Rwanda 68. United Kingdom
19. Senegal 69. Barbados
20. Sierra Leone 70. Canada
21. South Africa 71. Costa Rica
22. Sudan 72. Dominican Republic
23. Swaziland 73. El Salvador
24. Tanzania 74. Guatemala
25. Togo 75. Haiti
26. Tunisia 76. Honduras
27. Uganda 77. Jamaica
28. Zaire 78. Mexico
29. Zambia 79. Nicaragua
30. Zimbabwe 80. Panama
31. Bangladesh 81. Trinidad and Tobago
32. Burma 82. United States
33. Hong Kong 83. Argentina
34. India 84. Bolivia
35. Iran 85. Brazil
36. Israel 86. Chile
37. Japan 87. Colombia
38. Jordan 88. Ecuador
39. Korea 89. Guyana
40. Malaysia 90. Paraguay
41. Nepal 91. Peru
42. Pakistan 92. Uruguay
43. Philippines 93. Venezuela
44. Singapore 94. Australia
45. Sri Lanka 95. Fiji
46. Taiwan 96. New Zealand
47. Thailand 97. Papua New Guinea
48. Austria 98. Indonesia
49. Belgium
50. Cyprus
250 JOURNAL OF POLITICAL ECONOMY

References
Barro, Robert J. "Economic Growth in a Cross Section of Countries." QJ.E.
106 (May 1991): 407-43.
Barro, Robert J., and Sala-i-Martin, Xavier. "Convergence across States and
Regions." BrookingsPapers Econ. Activity, no. 1 (1991), pp. 107-82. (a)
. "Economic Growth." Manuscript. Cambridge, Mass.: Harvard Univ.,
1991. (b)
Blanchard, Olivier J., and Fischer, Stanley. Lectureson Macroeconomics.Cam-
bridge, Mass.: MIT Press, 1989.
Bureau of Economic Analysis. State Personal Income by State: 1929-1982.
Washington: Government Printing Office, 1986.
Cass, David. "Optimum Growth in an Aggregative Model of Capital Accumu-
lation." Rev. Econ. Studies 32 (July 1965): 233-40.
Chamley, Christophe. "The Welfare Cost of Capital Income Taxation in a
Growing Economy."J.P.E. 89 (June 1981): 468-96.
Cohen, Daniel, and Sachs, Jeffrey. "Growth and External Debt under Risk
of Debt Repudiation." European Econ. Rev. 30 (June 1986): 529-60.
Easterlin, Richard A. "Interregional Differences in Per Capita Income, Popu-
lation, and Total Income, 1840-1950." In Trends in the AmericanEconomy
in theNineteenth Century.Princeton, N.J.: Princeton Univ. Press (for NBER),
1960. (a)
. "Regional Growth of Income: Long Term Tendencies." In Population
Redistributionand Economic Growth, United States, 1870-1950, vol. 2, Analysis
of EconomicChange, edited by Simon Kuznets, Ann Ratner Miller, and Rich-
ard A. Easterlin. Philadelphia: American Philosophical Soc., 1960. (b)
Jorgenson, Dale W., and Yun, Kun-Young. "Tax Policy and Capital Alloca-
tion." Scandinavian J. Econ. 88, no. 2 (1986): 355-77.
. "Tax Reform and U.S. Economic Growth." J.P.E. 98, no. 5, pt. 2
(October 1990): S 151 -S 193.
King, Robert G., and Rebelo, Sergio T. "Transitional Dynamics and Eco-
nomic Growth in the Neoclassical Model." Working Paper no. 3185. Cam-
bridge, Mass.: NBER, November 1989.
Koopmans, Tjalling C. "On the Concept of Optimal Economic Growth." In
The Econometric Approach to Development Planning. Amsterdam: North-
Holland, 1965.
Maddison, Angus. "Growth and Slowdown in Advanced Capitalist Econo-
mies: Techniques of Quantitative Assessment."J. Econ. Literature25 (June
1987): 649-98.
Mulligan, Casey B., and Sala-i-Martin, Xavier. "Two-Capital-Goods Models
of Endogenous Growth: Steady State, Transitional Dynamics, and Empiri-
cal Implications." Manuscript. New Haven, Conn.: Yale Univ., June 1991.
Nelson, Richard R., and Phelps, Edmund S. "Investment in Humans, Tech-
nological Diffusion, and Economic Growth." A.E.R. Papers and Proc. 56
(May 1966): 69-75.
Ramsey, Frank P. "A Mathematical Theory of Saving." Econ.J. 38 (December
1928): 543-59.
Rebelo, Sergio. "Long-Run Policy Analysis and Long-Run Growth."J.P.E. 99
(June 1991): 500-521.
Renshaw, Vernon; Trott, E. A., Jr.; and Friedenberg, Howard L. "Gross
State Product by Industry, 1963-1986." SurveyCurrentBus. 68 (May 1988):
30-46.
CONVERGENCE 251

Sala-i-Martin, Xavier. "On Growth and States." Ph.D. dissertation, Harvard

Univ., 1990.
Sato, Kazuo. "On the Adjustment Time in Neo-classical Growth Models."
Rev. Econ. Studies 33 (July 1966): 263-68.
Solow, Robert M. "A Contribution to the Theory of Economic Growth."
QJ.E. 70 (February 1956): 65-94.
Summers, Robert, and Heston, Alan. "A New Set of International Compari-
sons of Real Product and Price Levels: Estimates for 130 Countries, 1950-
1985." Rev. Income and Wealth 34 (March 1988): 1-25.